FloorPlanBench
How well do computational methods generate valid building floor plans while trading off spatial efficiency, circulation, daylight, privacy and an embodied-carbon proxy? Seven classical methods, 28 benchmark programmes in four complexity tiers, 10 seeds each, one shared evaluation budget, one validator that no plan bypasses.
Three-bedroom bungalow (3b5p, ~100 m², L-plan) — balanced plan from the union Pareto front, found by CP search (seed 8). Inspect →
Headline results
| Method | Tasks solved | Median HV | Weight-robust wins |
|---|---|---|---|
| Random | 12% | 0.000 | 0% |
| Greedy | 92% | 0.319 | 5% |
| SA | 71% | 0.341 | 9% |
| Greedy→SA | 92% | 0.363 | 63% |
| NSGA-II | 44% | 0.196 | 6% |
| CP search | 80% | 0.292 | 18% |
| Analytic (grad.) | 41% | 0.000 | 0% |
Tasks solved = share of (task, seed) runs that produced at least one valid plan. HV = exact hypervolume of the 5 headline objectives, reference point 1.0001⁵, median over tasks of the per-task median over seeds. Weight-robust wins = share of 300 random Dirichlet weightings × tasks under which the method holds the best scalarised plan.
Research questions
- Which algorithms most reliably produce valid plans? → validity
- Which methods best explore Pareto trade-offs? → Pareto explorer
- How does performance degrade as constraints increase? → complexity scaling
- Do generative/ML-style methods outperform classical optimisation after validity is considered? → RQ4
- How sensitive are conclusions to objective weighting? → sensitivity
What makes this a benchmark rather than a plan generator
Integer geometry, exact validator
Every coordinate is an integer decimetre. Overlap, containment (including L/U notches), shared walls, door placement, connectivity and required adjacency are exact tests. Invalid plans never enter any leaderboard, front or gallery: the report re-validates every stored plan before writing.
Raw objectives, explicit maths
Nine objectives are kept raw: usable-area ratio, preferred-adjacency satisfaction, door-graph circulation distance, room compactness, a daylight-zone proxy, privacy depth, exterior exposure, accessibility clearances and a partition-length carbon proxy. Pareto fronts, hypervolume and 300-weight sensitivity replace any single score.
One budget, one door policy
All methods spend the same 50,000 evaluation-equivalents; partial work (candidate checks, gradient steps) is charged at measured rates. Doors and windows are a deterministic function of room rectangles, so methods compete on geometry only.